Part 1 of 3 in the series "AI on Business Data — Reliable, Not Just Plausible": Part 1: ChatGPT vs. oneAgent (this article) · Part 2: Correct SQL, Wrong Metric · Part 3: Query Your Data Warehouse with AI
"But we already have ChatGPT."
I hear this sentence regularly in conversations about oneAgent. And the question behind it is fair:
Why does a company need a specialized solution like oneAgent when it already uses ChatGPT, Claude, or Gemini?
The simple answer would be: because those tools can't access business data.
But that's no longer true.
General AI assistants can analyze files and access business systems through integrations. So the decisive question is no longer:
Can the AI access our data?
But rather:
Can it understand that data according to our business rules and evaluate it reliably?
This is exactly where the gap begins that we close with oneAgent.
Data access is not data understanding
Let's ask a seemingly simple question:
"What was our revenue last month?"
To a human, that sounds unambiguous. In the data model, however, it hides many decisions:
- Gross or net?
- Before or after discounts?
- Are returns deducted?
- Do cancelled orders count?
- Does the order, delivery, or invoice date apply?
- Which entities and currencies are included?
An AI can have access to every table it needs and still use a different revenue definition than Controlling or Power BI.
The result can be convincingly worded and technically correctly computed — and still be wrong in business terms. How exactly that happens is what we show in Part 2: Correct SQL, Wrong Metric.
That's why, in AI data analysis, it's not just important which language model is used.
More important is:
Which verified business logic is applied to the answer?
What general AI assistants do very well
ChatGPT, Claude, and Gemini are powerful all-round assistants. They can:
- write and summarize texts,
- analyze documents,
- gather information from different sources,
- develop ideas,
- evaluate files,
- and access work systems through integrations.
Their strength lies in breadth. They are meant to solve as many tasks as possible for as many people as possible.
oneAgent, by contrast, was built for a much more specific job:
Answering business questions in natural language based on structured company data.
A language model doesn't automatically know your rules
A general AI assistant knows language and context.
But it doesn't automatically know:
- your revenue definition,
- your permitted table relationships,
- your default filters,
- your roles and permissions,
- or the differences between similar-sounding metrics.
Of course, such rules can also be built around a general language model.
But then the company needs, among other things:
- documented data models,
- its own integrations,
- metric definitions,
- permission logic,
- validations,
- and a user interface.
At that point you're no longer simply using ChatGPT, Claude, or Gemini.
You're building your own data application on top of a language model.
This is exactly where the build-or-buy decision begins.
Which gap oneAgent closes
oneAgent is not meant to replace ChatGPT for writing emails, research, or programming.
oneAgent connects three layers:
- the question of a business user,
- the existing data platform,
- the verified logic behind the company's metrics.
What that looks like in practice on an existing data warehouse is covered in Part 3: Query Your Data Warehouse with AI.
This creates several key differences.
Metrics are not reinvented with every question
Revenue, margin, or return rate should not be reinterpreted with every request.
The business logic is defined centrally and used in a controlled way.
The language model understands the question. The metric logic computes the answer.
The data model gets business context
Table and column names are rarely enough.
A column called amount can mean invoice value, order value, or incoming payment. That's why oneAgent connects technical metadata with business knowledge:
- What does the table mean?
- Which columns are relevant?
- How is the data connected?
- Which relationships may be used?
Ambiguity triggers a clarifying question
A fast answer is not automatically a good answer.
If a question is unclear, the system should not simply pick one interpretation.
It should ask, for example:
"Do you mean revenue by order date or invoice date?"
Clarifying questions are not a sign of weakness. They prevent a convincingly worded answer from being based on the wrong interpretation.
Permissions apply in the chat, too
An AI agent must not show a user more data than that user is allowed to see.
Permissions must therefore be enforced all the way into the actual query — for entities, regions, locations, or sensitive data.
Results can be reused
Many questions repeat:
- How is revenue developing?
- Which cost centers are over budget?
- Which products are losing margin?
In oneAgent, relevant results can be saved and reused in dashboards.
A spontaneous question becomes a permanently usable analysis.
A concrete example
A managing director asks:
"Why did our revenue drop in May?"
For a reliable answer, the following questions need to be settled first:
- Which revenue definition applies?
- Which date logic belongs to it?
- How are cancellations and returns handled?
- Which comparison period makes sense?
- Which data is the user allowed to see?
- Is the question unambiguous?
Only then does it make sense to analyze which products, regions, or customer groups caused the decline.
The difference is not primarily in the worded explanation.
It lies in how the number before it came to be.
General AI assistant and oneAgent compared
| Question | General AI assistant | oneAgent |
|---|---|---|
| Main job | Universal assistant for many areas of work | Specialized analysis of structured business data |
| Data access | Via uploads, apps, integrations, or MCP | Via connected business data and existing data platforms |
| Metric logic | Must be provided individually or built additionally | Core part of the analysis |
| Table relationships | Can be inferred from context and documentation | Modeled by business meaning and used in a controlled way |
| Ambiguous questions | Depends on model, prompt, and configuration | Targeted clarifying questions are part of the analysis process |
| Permissions | Depends on product, integration, and configuration | Must be enforced down to the actual data query |
| Reuse | Answers often remain part of individual chats | Results can be saved and used in dashboards |
| Typical use | Research, documents, texts, and general tasks | Recurring and decision-relevant questions about business data |
The table doesn't mean general platforms fundamentally can't cover the right-hand column — it shows what each product puts at its center by default.
When a general AI assistant is enough
ChatGPT, Claude, or Gemini can be a good choice when:
- a single Excel or CSV file is being examined,
- it's about an exploratory analysis,
- no binding company metric is involved,
- documents need to be summarized,
- or a technical team deliberately builds its own solution.
We use general AI assistants ourselves every day. What to watch out for with sensitive data is covered in Entering company data into ChatGPT? and Entering customer data into ChatGPT.
So the question is not:
"Which tool is fundamentally better?"
But rather:
What task was it built for?
When a specialized solution makes sense
oneAgent becomes relevant when:
- current business data is queried regularly,
- multiple tables or data sources interact,
- metrics must be defined consistently company-wide,
- results influence business decisions,
- different users may see different data,
- and business users should work without SQL or DAX skills.
The more important a data question is, the more important control and traceability become. For a direct feature comparison, including against Excel, see oneAgent vs. Excel & ChatGPT.
Why we build oneAgent
When we started working more intensively on AI data analysis, one thing became clear quickly:
The language model alone is not the product.
The hard questions only begin afterwards:
- What does a metric mean?
- Which data belongs to it?
- Which relationships are correct?
- When must the system ask a clarifying question?
- How do we prevent plausible but wrong numbers?
- How do we give more people access to data without giving up governance?
That's why we're not building oneAgent as yet another general chatbot.
We're building a controlled access layer between the people in a company and the data logic that already exists.
Flexible like ChatGPT. Controlled like BI. Safe for business data.
Conclusion
ChatGPT, Claude, and Gemini can access business information.
But data access still doesn't answer:
- what revenue means,
- which calculation is approved,
- which relationships are correct,
- and which data a user is allowed to see.
The decisive gap lies between language and business logic.
Or put more simply:
ChatGPT, Claude, and Gemini understand your question. oneAgent additionally understands the rules by which your company computes the answer.
Keep reading
Part 2 of the series shows with concrete examples how SQL can run without a single error and still return the wrong metric. Continue to Part 2: Correct SQL, Wrong Metric.